Agentic Finance Graph

Home / Detections

Detections · found by our own detectors

What our detectors caught.

Every row below is a detection: something that happened at a time, with evidence behind it — a burst of registrations minted by one address inside an hour, an agent paying the same counterparty over and over, an endpoint that has never once answered, an identity that changed owner. A low score is not a detection and you will not find one here; that is a measurement, and it lives on the data page.

A handful of detections are also incidents: a harm, happening to somebody, recently. They are shown first and in red. Most days there are none, and saying so is the point.

Live incidents

5

Registration #529727 paid 0xf27313…b42f 647 times in a run, 85 of them inside ten minutes647
live incident2026-11-09 19:06 UTC#529727agent page
Registration #1131574 changed owner — the wallet we joined it to is no longer its wallet
live incident2026-11-08 19:14 UTC#1131574agent page
Registration #58977 changed owner — the wallet we joined it to is no longer its wallet
live incident2026-11-08 17:22 UTC#58977agent page
Registration #529727 paid 0xf27313…b42f 2460 times in a run, 99 of them inside ten minutes2,460
live incident2026-11-08 14:57 UTC#529727agent page
Registration #88292 paid 0xd334ab…17ed 365 times in a run, 184 of them inside ten minutes365
live incident2026-11-08 03:30 UTC#88292agent page
Registration #529727 paid 0xf27313…b42f 3090 times in a run, 100 of them inside ten minutes3,090
live incident2026-11-07 14:37 UTC#529727agent page
Detections logged 45,364 Detectors run hourly against measurements we already hold — no external calls, so this costs nothing to keep current. loading…
Last 7 days 55 Dated by when the event happened, not when we noticed it — a backfill does not make last week look busy. rolling 7 days
Major 0 Reserved for scale: a burst of 500+ registrations from one address in one hour, and the like. Severity is size, not harm. severity: major
Kinds firing 6 Detectors that have found something. Several more are armed and quiet, which is the correct state for a detector with nothing to report. distinct kinds

Each rule's record

A drain-shaped payment, a dead endpoint and a confirmed theft cost very different amounts when the rule is wrong, so every rule carries its own record beside its latest fire: how often it fired, how many fires we later re-derived, how many held, how many we withdrew, how many came next to an address-poisoning attempt, and how many are live. A blank cell means we hold no record that could fill it. JSON · definition

Detector record · as of 2026-11-10 06:54 UTC · poisoning watch from 2026-06-05
RuleFiredLive nowRe-derivedHeldWithdrawnNext to poisoningLatest fire
test then drain200——12026-11-06
retry storm6154717239478—2026-11-09
drain pattern
Bankrbot wallet theft, May 2026: live missed, replay fires
420422616142026-11-08
fleet sibling exposure
Unscored: shapes we log, not verified detections
1360———322026-10-29
mint burst
Unscored: shapes we log, not verified detections
3650———22026-11-02
god key unknown 7702
Unscored: shapes we log, not verified detections
00————never
uri timeout streak
Unscored: shapes we log, not verified detections
7190————2026-11-10
bind closed after nft transfer
Unscored: shapes we log, not verified detections
43,4672———32026-11-09
layering fan
Unscored: shapes we log, not verified detections
160————2026-10-28
address poisoning loss2—220—2026-09-25

Measured rules

Every rule above was run against synthetic data with known answers before it shipped: an adaptation of IBM's AMLSim that generates thousands of agent wallets with drains, retry storms, mint bursts and layering rings planted inside, then runs these very detectors over it (generator, validator and reports, Apache-2.0). Synthetic data measures the rules, not the chain: a number here says how a rule behaves on a shape, and nothing about how often the shape occurs in the real ledger. Weekly run of 11 October 2026, seed 20261010: 2,813 agents, 66,900 transfers, 189 planted patterns.

RulePlantedCaughtFalse alarmsPrecisionStatus
retry_storm v2 (run, ten-minute peak)1515, plus the 8 hour-straddling storms v1 missed0100%live since 11 Oct
mint_burst v2 (run, sixty-minute peak)660100%live since 11 Oct
drain_pattern v1 rule alone252511 (all one-off purchases)69.4%live; the balance reading below is now published beside each fire
drain_pattern with the balance test (took 80%+ of the wallet)25250100%evidence on every detection since 11 Oct; a filter after a month of side-by-side numbers
layering_fan (same-amount fan-out)660100%new kind, 11 Oct
uri_timeout_streak12120100%live
slow drain (a wallet emptied over 72 hours in several payments)1010486%not published: on the live ledger the test fires on ordinary hot wallets that run near zero
slow drain by the share of outflow the new payee took (no chain read)1010190.9%candidate, not live: on the live ledger it keeps 44 of 156 windows all time (5 of 25 in the last 30 days), which a human reads first; with its full conditions it kept 3 windows on 13 October that the 11 October reading did not pick, so it stays off until rule and reading agree (the 44 windows)

Also checked on the same run: the liveness ladder (2,813 of 2,813 agents at the level the published rule says), the pass-through rule (10 of 10 cases) and the address-poisoning matcher (12 of 12). The run repeats every Sunday; the table is updated at each release.

By kind · tap one to filter the log and read the rule

The log · newest first · tap a row for everything we hold on it

WhenKindSeveritySubjectCountWhat was seen
2026-11-10 06:24 UTCuri timeout streakinfo0x8004A1…86375Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:28637 has failed 5 probes and has never answered
2026-11-10 00:19 UTCuri timeout streakinfo0x8004A1…30205Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:23020 has failed 5 probes and has never answered
2026-11-10 00:18 UTCuri timeout streakinfo0x8004A1…06915Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:30691 has failed 5 probes and has never answered
2026-11-09 19:06 UTClive · retry stormnotable#529727647Registration #529727 paid 0xf27313…b42f 647 times in a run, 85 of them inside ten minutes
2026-11-09 18:12 UTCuri timeout streakinfo0x8004A1…57815Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:55781 has failed 5 probes and has never answered
2026-11-09 12:20 UTCuri timeout streakinfo0x8004A1…19025Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:1902 has failed 5 probes and has never answered
2026-11-09 03:14 UTCbind closed after nft transferinfo#1173933—Registration #1173933 changed owner — the wallet we joined it to is no longer its wallet
2026-11-09 02:58 UTCbind closed after nft transferinfo#1173894—Registration #1173894 changed owner — the wallet we joined it to is no longer its wallet
2026-11-08 23:13 UTCdrain patternnotable#58971—Registration #58971 sent 29x its previous largest payment to an address it had never paid, 100% of what the wallet held
2026-11-08 19:14 UTClive · bind closed after nft transfernotable#1131574—Registration #1131574 changed owner — the wallet we joined it to is no longer its wallet
2026-11-08 18:35 UTCdelegate changedinfo#56706—delegate changed
2026-11-08 18:18 UTCuri timeout streakinfo0x8004A1…88855Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:38885 has failed 5 probes and has never answered
2026-11-08 17:22 UTClive · bind closed after nft transfernotable#58977—Registration #58977 changed owner — the wallet we joined it to is no longer its wallet
2026-11-08 14:57 UTClive · retry stormnotable#5297272,460Registration #529727 paid 0xf27313…b42f 2460 times in a run, 99 of them inside ten minutes
2026-11-08 12:14 UTCuri timeout streakinfo0x8004A1…27515Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:52751 has failed 5 probes and has never answered
2026-11-08 06:24 UTCuri timeout streakinfo0x8004A1…09845Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:20984 has failed 5 probes and has never answered
2026-11-08 04:03 UTCbind closed after nft transferinfo#1171226—Registration #1171226 changed owner — the wallet we joined it to is no longer its wallet
2026-11-08 03:52 UTCbind closed after nft transferinfo#1171207—Registration #1171207 changed owner — the wallet we joined it to is no longer its wallet
2026-11-08 03:30 UTClive · retry stormnotable#88292365Registration #88292 paid 0xd334ab…17ed 365 times in a run, 184 of them inside ten minutes
2026-11-08 03:27 UTCbind closed after nft transferinfo#1171164—Registration #1171164 changed owner — the wallet we joined it to is no longer its wallet
2026-11-08 03:20 UTCbind closed after nft transferinfo#1171153—Registration #1171153 changed owner — the wallet we joined it to is no longer its wallet
2026-11-08 00:26 UTCuri timeout streakinfo0x8004A1…21225Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:52122 has failed 5 probes and has never answered
2026-11-08 00:25 UTCuri timeout streakinfo0x8004A1…01825Registration eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432:30182 has failed 5 probes and has never answered
2026-11-07 21:38 UTCdrain patternnotable#57191—Registration #57191 sent 84x its previous largest payment to an address it had never paid, 100% of what the wallet held
2026-11-07 14:37 UTClive · retry stormnotable#5297273,090Registration #529727 paid 0xf27313…b42f 3090 times in a run, 100 of them inside ten minutes

45,364 logged · 1,975 more with a Desk plan · JSON API · live incidents only · how the ranks work

Detection, incident, and the difference.

New, 11 October 2026: three corrections and a new kind. Retry storms and mint bursts are now found as runs with a peak: a storm is a run of payments to one payee with no gap over ten minutes and some ten-minute span holding fifty or more; a burst the same with sixty minutes and twenty-five. The calendar-hour buckets they replace missed every storm that crossed the hour mark: 10 storm detections were logged where the ledger holds 45, from 7 payer–payee pairs where there are 28, and 83 bursting addresses where there are 88. Rows logged before 11 October stay in the log as history. Every drain detection now carries a balance reading, how much of what the wallet held the payment took, read from the chain at the block before: on the 47 drain-shaped payments since February, 39 took 80 to 100 percent. It is evidence beside the rule, not yet a filter. And layering_fan is new: one agent paying four or more others the same exact amount within days, the one layering shape the chain has shown (ten all time, each a fleet being funded; no cycle and no scatter-gather in the 314,826 agent-to-agent payments walked on 10 October). A shape, never an incident. Lists are now capped with a warning instead of silently. Definitions: retry_storm.v2, mint_burst.v2, drain_pattern.v2, layering_fan.v1.

New, 24 September 2026: drain_pattern. A payment at least five times the agent’s previous largest, at least $250, to an address it had never paid, after at least three earlier payments. It is the shape a drained agent wallet leaves, and also the shape of a legitimate one-off purchase, so it is always a detection for a human to read, never an incident, and the agent is never named publicly. The reasoning →

A detection is an occurrence: it happened at a time, we can point at the evidence, and a reader could in principle have watched it happen. A detector that fired on a state would turn this page into a scoreboard, and a scoreboard of low scores is worth nothing to anybody. "This agent has never answered a probe" sits on the edge — it is a state — so it only becomes a detection after a streak of failures with no success ever recorded, which makes it an event about our own repeated attempts rather than a verdict on the agent.

An incident is a detection that is also a harm, and recent. Three kinds qualify: a retry storm, where an agent pays the same counterparty over and over and somebody is paying the fees for every attempt; an identity that changed owner under a binding people rely on, so the history they trust now belongs to somebody else; and signing delegated to an implementation nobody has vetted. Each counts as live for 72 hours, then stays in the log as a detection.

A mint burst is never an incident, however large. One address minting hundreds of registrations in an hour is exactly what a platform issuing identities for its users looks like. It is worth logging because it changes what a registration count means — the headline is measuring a script, not an ecosystem — but it harms nobody. This page called it a sybil burst until September 2026, and that was wrong: sybil asserts intent, and our own definition says the opposite.

Detectors that are armed and currently silent still matter. god_key_unknown_7702 fires when a wallet delegates to an implementation outside a curated allowlist, and that allowlist ships empty on purpose — until a human has read a contract and can say why it belongs there, calling an unknown implementation a god key would be a guess, and this site does not publish guesses. It reports nothing today, and that is the honest answer rather than a broken one.

Get these as alerts.

Everything live, and a morning digest: the Telegram group t.me/AgenticFinanceGraph posts each live incident as it is logged and a digest every morning. The digest is also how you know the feed is alive on a quiet day.

Your own agents: message @AgenticGraphBot and send /watch 57657 (an ERC-8004 number on Base, its slug, or the bound wallet). Free, three agents per chat. The bot tells you when our detectors log something on one of them, and /agent 57657, /top and /detections answer on demand. Every agent page has a Watch on Telegram button that opens the bot with that agent already watched. /stop deletes the chat's id and its list.

I joined the group but see no alerts: alerts post only when a detector logs a live incident, and the morning digest proves the feed is running. The bot did not answer /watch: send the agent's number, not its name. Email alerts are next. A chat watches 1 agent free, or 20 or 60 when it is linked to an Operator or Team Desk (/link).